The Henry Ford moment for software

Henry Ford did not invent the car. Cars existed, hand-built by craftsmen for buyers who could afford craft. He did not invent the assembly line either. Ransom Olds ran one before him. What Ford built at Highland Park was the moving assembly line and the production system around it: consistent output at scale, governed by process instead of individual skill. That system was the invention. Early factory output was often no better than the artisan's, and sometimes worse. But the line ran faster than any workshop could match, and speed with consistency beat craft.

The artifact arrives first and takes the headlines. The production system arrives second and takes the economy.

For decades, software has been a trade of individual hands: an experienced engineer can often guess the author of a file from its style alone. The whole quality model, code review, pair programming, seniority ladders, rests on individual judgment. That model has one assumption buried in it: the code arrives at a speed a human can judge.

AI now writes code faster than any organization can review it, and with machine variance: the same prompt that produces a sound system today can produce a different one tomorrow. Review was built for a trickle of human-paced changes. It now faces a flood.

The flood is not clean. Veracode's 2025 tests across more than 100 models found AI-generated code introduced a security flaw in 45%[5] of tasks. In March 2026, DryRun Security, itself an application-security vendor, reported worse from a small sample: 26 of 30 AI-agent pull requests, 87%[4], carried at least one vulnerability.

You do not need AI to see what one ungoverned change costs. In March 2026, Business Insider reported from internal Amazon documents that a single configuration change made outside change management took Amazon.com down for about six hours and cost 6.3 million lost orders[15]. The response described in the same reporting leaned on more manual sign-offs. The sign-off itself is not the mistake. Regulators require a named human accountable for every change, and that stays. What fails is asking that human to assemble the evidence by hand while change arrives at machine speed. Keep the accountability. Automate the evidence.

In April 2026, a16z published adoption data: 29% of the Fortune 500 pay for products from the AI startups it tracks, and coding tools are the standout category. Kimberly Tan's thesis: "Code is upstream of all other applications because it's the core building block for any piece of software, so AI's accelerating impact on code should accelerate every other domain."

Half right. Code generation got cheaper, even if the multiplier is contested. The 10 to 20x figure is what a16z says it hears from its portfolio companies' best engineers. Independent measurement runs the other way in places: in METR's 2025 randomized trial, experienced developers were about 19% slower with AI assistance on mature codebases. Wherever the truth sits, generation is no longer the constraint that decides enterprise delivery.

Inside a regulated enterprise, code is not the top of the stream. Upstream of code sit the specification, the governance review, the compliance evidence, and the audit trail. These are not abstractions. DORA requires financial entities to keep auditable records of every ICT change. NIS2 puts similar duties on critical infrastructure. That is where most of the delay lives. Compress only the build and you get a clever weekend. Compress what sits upstream of the build and the same speed ships something a regulator can audit.

The Henry Ford moment for software is not a better model. Models keep improving, and every improvement makes the output faster and cheaper without making it governed. The moment is the assembly line itself: the production system that takes machine-speed output and makes it auditable in the sense a regulator means it, every change traceable to the intent that ordered it and the approval that let it ship.

A real one looks stranger than demo culture suggests. It is deliberately slow, hours and sometimes days, because specification work, policy checks, and human approvals run before and alongside the build rather than after it. Intake comes before generation: the system draws intent and policy out of the organization until it knows enough to build. In Taiga's case the conversation runs in reverse. You do not prompt the factory. The factory prompts you.

Agents improvise. Factories repeat. Repeat means the deliverable, never the bytes, which nobody can honestly promise. Taiga's factory, for instance, delivers the same nine artifacts[8] against the same published 22-check bar[7] on every run.

A published bar does not cancel the vulnerability numbers above. It changes the question. Instead of asking whether a human reviewer caught everything in an unbounded stream, you ask whether the published checks cover the failure classes the studies found. Gate coverage becomes the critical question, and unlike reviewer attention, it is a question you can inspect.

The carpenters did not disappear when the factories came. Furniture still needed people who understood wood, and those people designed and ran the production lines. The ones who survived were not the finest carvers. They were the ones who understood the factory.

Developers get the same sorting. The work moves upstream: define the intent precisely, write the policy the machine enforces, judge the run that passed every check and still missed the point, and send it back. Coding assistants keep making individual developers faster, and the factory governs what the organization ships. Complement, not competitor.

The production system for machine-speed software is the AI software factory: business intent in, governed production-grade software out, audit trail attached.

Renting a factory raises a question Ford never faced. The vendor becomes a critical ICT third party, the kind DORA names, and a regulated buyer should treat it as one. Ask whether the checks are published, whether the code lives in your own repositories, and whether the artifacts stay yours if you leave. A factory that fails its own audit is just a new dependency. The product page shows how Taiga answers today.

The car existed before Ford. The factory did not. The AI exists now, and this time the factory does too.